Artificial Intelligence Nanodegree¶
Convolutional Neural Networks¶
Project: Write an Algorithm for a Dog Identification App¶
Solution by Carsten Isert, Sept. 2017¶
In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!
Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the iPython Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.
The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this IPython notebook.
Why We're Here¶
In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).
In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!
The Road Ahead¶
We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.
- Step 0: Import Datasets
- Step 1: Detect Humans
- Step 2: Detect Dogs
- Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
- Step 4: Use a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 6: Write your Algorithm
- Step 7: Test Your Algorithm
Step 0: Import Datasets¶
Import Dog Dataset¶
In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:
train_files,valid_files,test_files- numpy arrays containing file paths to imagestrain_targets,valid_targets,test_targets- numpy arrays containing onehot-encoded classification labelsdog_names- list of string-valued dog breed names for translating labels
from sklearn.datasets import load_files
from keras.utils import np_utils
import numpy as np
from glob import glob
# define function to load train, test, and validation datasets
def load_dataset(path):
data = load_files(path)
dog_files = np.array(data['filenames'])
dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
return dog_files, dog_targets
# load train, test, and validation datasets
train_files, train_targets = load_dataset('dogImages/train')
valid_files, valid_targets = load_dataset('dogImages/valid')
test_files, test_targets = load_dataset('dogImages/test')
# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]
# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
Import Human Dataset¶
In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.
import random
random.seed(8675309)
# load filenames in shuffled human dataset
human_files = np.array(glob("lfw/*/*"))
random.shuffle(human_files)
# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
Step 1: Detect Humans¶
We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.
In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
import cv2
import matplotlib.pyplot as plt
%matplotlib inline
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
img = cv2.imread(human_files[3])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
Write a Human Face Detector¶
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
(IMPLEMENTATION) Assess the Human Face Detector¶
Question 1: Use the code cell below to test the performance of the face_detector function.
- What percentage of the first 100 images in
human_fileshave a detected human face? - What percentage of the first 100 images in
dog_fileshave a detected human face?
Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer: The performance on the human dataset is acceptable, but for the dogs it is quite bad with 11%.
- Percentage of detected faces in human dataset: 0.98 - 0.99
- Percentage of detected faces in dog dataset: 0.11
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
# Do NOT modify the code above this line.
number_of_detected_faces = 0
for file in human_files_short:
if face_detector(file):
number_of_detected_faces += 1
percentage_of_detected_faces = number_of_detected_faces / 100
print("Percentage of detected faces in human dataset: ", percentage_of_detected_faces)
number_of_detected_faces = 0
for file in dog_files_short:
if face_detector(file):
number_of_detected_faces += 1
percentage_of_detected_faces = number_of_detected_faces / 100
print("Percentage of detected faces in dog dataset: ", percentage_of_detected_faces)
Question 2: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). In your opinion, is this a reasonable expectation to pose on the user? If not, can you think of a way to detect humans in images that does not necessitate an image with a clearly presented face?
Answer: If we design the UI in a way that clearly communicates this to the user, I think this is a totally reasonable design choice. We are talking about a fun app anyway and people are of course free to also try this with photos of their cats or horses or cars and find out what dog breed they resemble. This might just be rubbish, but it also might be a lot of fun.
One approach to solve this, is to make a too stage approach. The first stage will try to find the location of faces in an image and the corresponding bounding box, and the second stage will take this area and do the analysis. This approach has been used e.g. at Baidu to make a face detection framework for allowing employees to enter restricted areas based on their face id.
We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on each of the datasets.
## (Optional) TODO: Report the performance of another
## face detection algorithm on the LFW dataset
### Feel free to use as many code cells as needed.
# Ideas for later:
# 1. Use the dog and human datasets to train a CNN that can distinguish between the humans and the dogs for this specific dataset.
# 2. Look for some existing algorithms that can be used (in TF or Keras). A first search didn't show anything promising.
Step 2: Detect Dogs¶
In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.
from keras.applications.resnet50 import ResNet50
# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')
Pre-process the Data¶
When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape
$$ (\text{nb_samples}, \text{rows}, \text{columns}, \text{channels}), $$where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.
The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape
The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape
Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in your dataset!
from keras.preprocessing import image
from tqdm import tqdm
# Note: modified these to functions, so that we can later also read the inception tensors which
# have a different format
def path_to_tensor(img_path, width=224, height=224):
# loads RGB image as PIL.Image.Image type
img = image.load_img(img_path, target_size=(width, height))
# convert PIL.Image.Image type to 3D tensor with shape (width, heigth, 3)
x = image.img_to_array(img)
# convert 3D tensor to 4D tensor with shape (1, width, height, 3) and return 4D tensor
return np.expand_dims(x, axis=0)
def paths_to_tensor(img_paths, width=224, height=224):
list_of_tensors = [path_to_tensor(img_path, width, height) for img_path in tqdm(img_paths)]
return np.vstack(list_of_tensors)
Making Predictions with ResNet-50¶
Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. If you're curious, you can check the code for preprocess_input here.
Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.
By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.
from keras.applications.resnet50 import preprocess_input, decode_predictions
def ResNet50_predict_labels(img_path):
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
Write a Dog Detector¶
While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).
We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
prediction = ResNet50_predict_labels(img_path)
return ((prediction <= 268) & (prediction >= 151))
(IMPLEMENTATION) Assess the Dog Detector¶
Question 3: Use the code cell below to test the performance of your dog_detector function.
- What percentage of the images in
human_files_shorthave a detected dog? - What percentage of the images in
dog_files_shorthave a detected dog?
Answer:
- Percentage of detected dogs in human dataset: 0.00 - 0.01
- Percentage of detected dogs in dog dataset: 1.0
number_of_detected_dogs = 0
for file in human_files_short:
if dog_detector(file):
number_of_detected_dogs += 1
percentage_of_detected_dogs = number_of_detected_dogs / 100
print("Percentage of detected dogs in human dataset: ", percentage_of_detected_dogs)
number_of_detected_dogs = 0
for file in dog_files_short:
if dog_detector(file):
number_of_detected_dogs += 1
percentage_of_detected_dogs = number_of_detected_dogs / 100
print("Percentage of detected dogs in dog dataset: ", percentage_of_detected_dogs)
Step 3: Create a CNN to Classify Dog Breeds (from Scratch)¶
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 1%. In Step 5 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
Be careful with adding too many trainable layers! More parameters means longer training, which means you are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; you can extrapolate this estimate to figure out how long it will take for your algorithm to train.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
 |
 |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
 |
 |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
 |
 |
 |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imbalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!
Pre-process the Data¶
We rescale the images by dividing every pixel in every image by 255.
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
model.summary()
We have imported some Python modules to get you started, but feel free to import as many modules as you need. If you end up getting stuck, here's a hint that specifies a model that trains relatively fast on CPU and attains >1% test accuracy in 5 epochs:
Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. If you chose to use the hinted architecture above, describe why you think that CNN architecture should work well for the image classification task.
Answer: My approach was to use the model from the cifar-10 classification as a baseline. Once I checked the number of parameters it was very huge. This was due to the size of the images being 224 x 224 instead of 32 x 32. So I reduced the number of parameters to enable faster training and avoid overfitting by increasing the sizes of the max pooling layers from 2 to 6. This helps to reduce the size of the layers and therefore the number of parameters. I also reduced the number of filters to 16 and 32 again to reduce the number of parameters. The dropouts were kept. The final step was to massively reduce the nodes in the last fully connected layer. I took a hint from the architecture above. It probably makes sense to have at least one neuron representing one dog breed. So I chose the number of nodes to 133. Additionally, I changed the optimizer to Adam, as in my experience, Adam works better as RMSProp, as it combines the methods from Momentum and RMSProp. The test accuracy after 10 epochs was 6.1005%. The validation accuracy in the last step was around 6.8% or even higher. That means that we might already be in some state of overfitting, but it is hard to tell at this stage. Later on, I had run the whole notebook and loaded the weihgts, but I had to run it again, so this way, I ran an additional 10 epochs and reached a accuracy of 10.6459%, so overfitting was not so severe at all. As I exceeded the requirements of 1% by far and training takes a lot of time, I didn't push much further. But I think there is a lot of room for improvement, which of course is shown in the later exercises.
I also compared my results to the excellent notebook from Jay Madhava at https://github.com/madhavajay/nd889/blob/master/2_deep_learning/1_dog_breed_classifier/dog_app.ipynb and found out that my network performs better without data augmentation. So I did not work further on this.
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
model = Sequential()
model.add(Conv2D(filters=16, kernel_size=3, padding='same', activation='relu',
input_shape=(224, 224, 3)))
model.add(Conv2D(filters=16, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=6))
model.add(Dropout(0.25))
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu'))
model.add(Conv2D(filters=32, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=6))
model.add(Dropout(0.3))
model.add(Flatten())
model.add(Dense(133, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(133, activation='softmax'))
model.summary()
Compile the Model¶
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
(IMPLEMENTATION) Train the Model¶
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
# helpers for training CNNs taken from
# https://github.com/madhavajay/nd889/blob/master/2_deep_learning/1_dog_breed_classifier/dog_app.ipynb
# many thanks to Jay for these helper functions
import keras
import timeit
# graph the history of model.fit
def show_history_graph(history):
# summarize history for accuracy
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
# summarize history for loss
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
# callback to show the total time taken during training and for each epoch
class EpochTimer(keras.callbacks.Callback):
train_start = 0
train_end = 0
epoch_start = 0
epoch_end = 0
def get_time(self):
return timeit.default_timer()
def on_train_begin(self, logs={}):
self.train_start = self.get_time()
def on_train_end(self, logs={}):
self.train_end = self.get_time()
print('Training took {} seconds'.format(self.train_end - self.train_start))
def on_epoch_begin(self, epoch, logs={}):
self.epoch_start = self.get_time()
def on_epoch_end(self, epoch, logs={}):
self.epoch_end = self.get_time()
print('Epoch {} took {} seconds'.format(epoch, self.epoch_end - self.epoch_start))
from keras.callbacks import ModelCheckpoint
epochs = 10
epochtimer = EpochTimer()
### Do NOT modify the code below this line.
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.from_scratch.hdf5',
verbose=1, save_best_only=True)
hist = model.fit(train_tensors, train_targets,
validation_data=(valid_tensors, valid_targets),
epochs=epochs, batch_size=20, callbacks=[checkpointer, epochtimer], verbose=1)
show_history_graph(hist)
Load the Model with the Best Validation Loss¶
model.load_weights('saved_models/weights.best.from_scratch.hdf5')
Test the Model¶
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 1%.
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0))) for tensor in test_tensors]
# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']
Model Architecture¶
The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.
VGG16_model = Sequential()
VGG16_model.add(GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(Dense(133, activation='softmax'))
VGG16_model.summary()
Compile the Model¶
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
Train the Model¶
epochtimer = EpochTimer()
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5',
verbose=1, save_best_only=True)
hist = VGG16_model.fit(train_VGG16, train_targets,
validation_data=(valid_VGG16, valid_targets),
epochs=20, batch_size=20, callbacks=[checkpointer, epochtimer], verbose=1)
show_history_graph(hist)
Load the Model with the Best Validation Loss¶
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')
Test the Model¶
Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0))) for feature in test_VGG16]
# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Predict Dog Breed with the Model¶
from extract_bottleneck_features import *
def VGG16_predict_breed(img_path):
# extract bottleneck features
bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = VGG16_model.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)¶
You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.
In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, you must use the bottleneck features from a different pre-trained model. To make things easier for you, we have pre-computed the features for all of the networks that are currently available in Keras:
- VGG-19 bottleneck features
- ResNet-50 bottleneck features
- Inception bottleneck features
- Xception bottleneck features
The files are encoded as such:
Dog{network}Data.npz
where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.
(IMPLEMENTATION) Obtain Bottleneck Features¶
In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:
bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']resnet_bottleneck_features = np.load('bottleneck_features/DogResnet50Data.npz')
train_resnet50 = resnet_bottleneck_features['train']
valid_resnet50 = resnet_bottleneck_features['valid']
test_resnet50 = resnet_bottleneck_features['test']
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
<your model's name>.summary()
Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.
Answer: I basically just did copy and paste from the VGG16 model, but renamed everything to Resnet50. This resulted in a performance / accuracy of 83.4928%, which far exceeds the requirements. It was interesting to observe that the model converged quite early (epoch 11) and then overfitted severely. This suggests to use additional methods to reduce this overfitting with dropouts, data augmentation as the dataset is quite small and a more complex model at the end, maybe with several fully conneted layers resembling the original structure of ResNet.
Some of these experiments were done (see section 8 for results and comments).
During these experiements, I also checked out to change the batch size to 64 and got test accuracy of 84.6890%. However, these results are not consistent, in the last run, the accuracy for these settings were only 82.7751%.
Resnet50_model = Sequential()
Resnet50_model.add(GlobalAveragePooling2D(input_shape=train_resnet50.shape[1:]))
Resnet50_model.add(Dense(133, activation='softmax'))
Resnet50_model.summary()
(IMPLEMENTATION) Compile the Model¶
Resnet50_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
(IMPLEMENTATION) Train the Model¶
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.RESNET50.hdf5',
verbose=1, save_best_only=True)
epochtimer = EpochTimer()
hist = Resnet50_model.fit(train_resnet50, train_targets,
validation_data=(valid_resnet50, valid_targets),
epochs=20, batch_size=64, callbacks=[checkpointer, epochtimer], verbose=1)
show_history_graph(hist)
(IMPLEMENTATION) Load the Model with the Best Validation Loss¶
Resnet50_model.load_weights('saved_models/weights.best.RESNET50.hdf5')
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 60%.
# get index of predicted dog breed for each image in test set
Resnet50_predictions = [np.argmax(Resnet50_model.predict(np.expand_dims(feature, axis=0))) for feature in test_resnet50]
# report test accuracy
test_accuracy = 100*np.sum(np.array(Resnet50_predictions)==np.argmax(test_targets, axis=1))/len(Resnet50_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
# Display some of the result images with correct or incorrect classification
fig = plt.figure(figsize=(20, 12))
for i, idx in enumerate(np.random.choice(test_tensors.shape[0], size=32, replace=False)):
ax = fig.add_subplot(4, 8, i + 1, xticks=[], yticks=[])
ax.imshow(np.squeeze(test_tensors[idx]))
true_idx = np.argmax(test_targets[idx])
pred_idx = Resnet50_predictions[idx]
ax.set_title("{}\n({})".format(dog_names[pred_idx], dog_names[true_idx]),
color=("green" if pred_idx == true_idx else "red"))
(IMPLEMENTATION) Predict Dog Breed with the Model¶
Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.
Similar to the analogous function in Step 5, your function should have three steps:
- Extract the bottleneck features corresponding to the chosen CNN model.
- Supply the bottleneck features as input to the model to return the predicted vector. Note that the argmax of this prediction vector gives the index of the predicted dog breed.
- Use the
dog_namesarray defined in Step 0 of this notebook to return the corresponding breed.
The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to your chosen CNN architecture, you need to use the function
extract_{network}
where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.
### Function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.
from extract_bottleneck_features import *
def Resnet50_predict_breed(img_path):
# extract bottleneck features
bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = Resnet50_model.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
Step 6: Write your Algorithm¶
Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
- if a dog is detected in the image, return the predicted breed.
- if a human is detected in the image, return the resembling dog breed.
- if neither is detected in the image, provide output that indicates an error.
You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 5 to predict dog breed.
Some sample output for our algorithm is provided below, but feel free to design your own user experience!
(IMPLEMENTATION) Write your Algorithm¶
import os
import fnmatch
# Helper function to be able to display a matching image for the humans so that we know immediately
# how the dog breed looks like, we search for the first image in the training data with the correct name.
def find_dog_img(pattern, path):
for root, dirs, files in os.walk(path):
for name in files:
if fnmatch.fnmatch(name, pattern):
return os.path.join(root, name)
import matplotlib.image as mpimg
def show_image(img_path):
img = mpimg.imread(img_path)
fig = plt.figure()
plt.subplot()
plt.imshow(img)
plt.axis('off')
plt.plot()
def dog_breed_or_similarity(img_path, show_images=False):
if show_images == True: show_image(img_path)
if dog_detector(img_path):
print("Such a nice dog, I would think you are a ...")
result = Resnet50_predict_breed(img_path)
print(result)
sample_dog_img = find_dog_img(result + '*', 'dogImages/train')
if show_images == True: show_image(sample_dog_img)
return result
if face_detector(img_path):
print("Hello Human, you look like a...")
result = Resnet50_predict_breed(img_path)
print(result)
sample_dog_img = find_dog_img(result + '*', 'dogImages/train')
if show_images == True: show_image(sample_dog_img)
return result
print("Neither dog or human. Make more effort!")
return "Error"
Step 7: Test Your Algorithm¶
In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?
(IMPLEMENTATION) Test Your Algorithm on Sample Images!¶
Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.
Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.
Answer: As this week is election in Germany, I took pictures from the major German politicians and ran them through this algorithm. Additionally, I took some human samples from the lfw folder, some private pictures and the pictures of dogs in the images directory and also the remaining images in this folder (network architecture). Overall, this was a really fun experience and the results are quite convincing for such a small project.
The following points of improvement could be done / have been done:
- Use data augmentation to generate more sample data as the dataset is quite small. Initially, I thought this could be easily done with the method learned in class, however, the bottleneck features are not images and so another architecture has to be used. Will be explored further. Results from other students show that the accuracy gain might not be too big. We need to use a full pre-trained network to feed in the augmented images. This also gives the opportunity for finetuning. See section 9.
- Use a more complex model at the end of the ResNet: See section 8.
- Check out another model. Reports on Slack mention that Inception achieves higher performance. As I have just read the ResNet paper, I wanted to use this one for this task. See section 8 for InceptionV3 an Xception results.
- Additionally, the human detection should be improved. One picture of a human was not classified correctly. (not addressed)
- It would be nice for the output of the algorithm to show the input image and a sample of the matching dog breed. This is implemented and can be run with the images available, but the images are not checked in due to potential copyright issues.
# Image outputs need to be run online and are not checked in to github
# due to potential copyright
input_directory = "/home/aind2/AIND-Dog/images"
list_of_files = os.listdir(input_directory)
file1 = 'carsten.png'
img_file = input_directory + '/' + file1
print("Working on file... ", img_file)
dog_breed_or_similarity(img_file, True)
# Image outputs need to be run online and are not checked in to github
# due to potential copyright
input_directory = "/home/aind2/AIND-Dog/images"
list_of_files = os.listdir(input_directory)
file1 = 'Angela_Merkel.jpg'
img_file = input_directory + '/' + file1
print("Working on file... ", img_file)
dog_breed_or_similarity(img_file, True)
import os
input_directory = "/home/aind2/AIND-Dog/images"
list_of_files = os.listdir(input_directory)
for file_name in list_of_files:
img_file = input_directory + '/' + file_name
print("Working on file... ", img_file)
# Somehow it looks like the plotting doesn't work easily in the loop
# So I would need to generate a nice table in the future...
# But it would be more reasonable to put this into an app anyway.
dog_breed_or_similarity(img_file)
print("-------------------------")
Step 8: Extensions¶
Evaluate two other models: Inception and Xception and play around with the top layers.
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
# Build Inception model
inception_bottleneck_features = np.load('bottleneck_features/DogInceptionV3Data.npz')
train_inception = inception_bottleneck_features['train']
valid_inception = inception_bottleneck_features['valid']
test_inception = inception_bottleneck_features['test']
Inception_model = Sequential()
Inception_model.add(GlobalAveragePooling2D(input_shape=train_inception.shape[1:]))
#Inception_model.add(Dense(64, activation='relu'))
#Inception_model.add(Dropout(0.3))
# Check out again...
#Inception_model.add(Flatten(input_shape=train_data.shape[1:]))
#Inception_model.add(Dense(256, activation='relu'))
#Inception_model.add(Dropout(0.5))
Inception_model.add(Dense(133, activation='softmax'))
Inception_model.summary()
Inception_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
from keras.callbacks import ModelCheckpoint
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.INCEPTION.hdf5',
verbose=1, save_best_only=True)
epochtimer = EpochTimer()
hist = Inception_model.fit(train_inception, train_targets,
validation_data=(valid_inception, valid_targets),
epochs=25, batch_size=32, callbacks=[checkpointer, epochtimer], verbose=1)
show_history_graph(hist)
Testing¶
Inception_model.load_weights('saved_models/weights.best.INCEPTION.hdf5')
# get index of predicted dog breed for each image in test set
Inception_predictions = [np.argmax(Inception_model.predict(np.expand_dims(feature, axis=0))) for feature in test_inception]
# report test accuracy
test_accuracy = 100*np.sum(np.array(Inception_predictions)==np.argmax(test_targets, axis=1))/len(Inception_predictions)
print('Test accuracy Inception: %.4f%%' % test_accuracy)
XCeption¶
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense, LeakyReLU
from keras.models import Sequential
# Build Xception model
xception_bottleneck_features = np.load('bottleneck_features/DogXceptionData.npz')
train_xception = xception_bottleneck_features['train']
valid_xception = xception_bottleneck_features['valid']
test_xception = xception_bottleneck_features['test']
Xception_model = Sequential()
Xception_model.add(GlobalAveragePooling2D(input_shape=train_inception.shape[1:]))
#Xception_model.add(Dense(64, activation='relu'))
#Xception_model.add(Dropout(0.3))
Xception_model.add(LeakyReLU(alpha=0.3))
Xception_model.add(Dense(512, activation='relu'))
Xception_model.add(LeakyReLU(alpha=0.3))
Xception_model.add(Dense(133, activation='softmax'))
Xception_model.summary()
Xception_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
from keras.callbacks import ModelCheckpoint
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.XCEPTION.hdf5',
verbose=1, save_best_only=True)
epochtimer = EpochTimer()
hist = Xception_model.fit(train_xception, train_targets,
validation_data=(valid_xception, valid_targets),
epochs=25, batch_size=32, callbacks=[checkpointer, epochtimer], verbose=1)
show_history_graph(hist)
Xception_model.load_weights('saved_models/weights.best.XCEPTION.hdf5')
# get index of predicted dog breed for each image in test set
Xception_predictions = [np.argmax(Xception_model.predict(np.expand_dims(feature, axis=0))) for feature in test_xception]
# report test accuracy
test_accuracy = 100*np.sum(np.array(Xception_predictions)==np.argmax(test_targets, axis=1))/len(Xception_predictions)
print('Test accuracy xception: %.4f%%' % test_accuracy)
fig = plt.figure(figsize=(20, 12))
for i, idx in enumerate(np.random.choice(test_tensors.shape[0], size=32, replace=False)):
ax = fig.add_subplot(4, 8, i + 1, xticks=[], yticks=[])
ax.imshow(np.squeeze(test_tensors[idx]))
true_idx = np.argmax(test_targets[idx])
pred_idx = Xception_predictions[idx]
ax.set_title("{}\n({})".format(dog_names[pred_idx], dog_names[true_idx]),
color=("green" if pred_idx == true_idx else "red"))
Results¶
The first extension experiment was with an exact copy of the previous ResNet architecture, just using the InceptionV3 bottleneck network upfront. This resulted in a test accuracy of: 82.7751%, i.e. worse than the Resnet50. The optimum result was already reached after 8 epochs, meaning that the architecture again did serious overfitting. In follow up experiments, this accuracy changed sometimes. One result was 80.3828%.
The second extension experiment was changing the network architecture. I tried, not to use the GlobalAveragePooling layer, but instead flatten the data and then connect another dense layer. Although the number of parameters were not so high (around 500,000), this model ran into the problem of resource allocation and could not be solved.
The third experiment was keeping the GlobalAveragePooling, but add another dense layer with dropouts. This resulted in an accuracy of 80.8612%. Again worse than before.
I tested the XCeption network with a top network using two LeakyReLus and a fully connected layer with 512 nodes. This resulted in an accuracy of 84.0909%, which is comparable to the ResNet50 performance, but involves a much more complicated network and longer training times. As can be seen in the figures, this network reached its accurcy peak for validation quite early and overfitted afterwards.
TODO: Playing around with batch size and optimizer.
9. Step Data Augmentation with Finetuning¶
As it is not possible to use data augmentation with the bottleneck models, the whole network has to be used.
As mentioned above I had a look at this tutorial Keras. However, it was not possible for me to build the network based on this information. So, I did a research on the Slack channels and found the excellent notebook and app of Jay Madhava. Most parts of the following code are taken from there, but with my own modifications and experiments to learn the transfer and fine tuning.
# InceptionV3 uses 299 x 299 images instead of the 224 x 224 above
# So we need to generate another set of input tensors.
from keras.applications.inception_v3 import preprocess_input
from tqdm import tqdm
def paths_to_inception_tensor(img_paths, width=299, height=299):
list_of_tensors = [preprocess_input(path_to_tensor(img_path, width, height)) for img_path in tqdm(img_paths)]
return np.vstack(list_of_tensors)
inception_test_tensors = paths_to_inception_tensor(test_files)
from keras.applications.inception_v3 import InceptionV3, preprocess_input
from keras.preprocessing.image import ImageDataGenerator
img_width, img_height = 299, 299
batch_size = 100
num_classes = 133
train_dir = 'dogImages/train'
valid_dir = 'dogImages/valid'
# Data augmentation basically in the same way it was introduced with the cifar-10 dataset
# Here we also use shear and zoom as the images are larger.
train_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True
)
# Same operation on the validation set
validation_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True
)
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
)
validation_generator = validation_datagen.flow_from_directory(
valid_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
)
from keras.models import Model
import math
# Load Inception model without the top layers
base_model = InceptionV3(weights='imagenet', include_top=False)
# Freeze all layers so that the weights cannot be updated
for layer in base_model.layers:
layer.trainable = False
# Construct our own model on top of Inception network
# First use same architecture as above with just GlobalAveragePooling and a dense layer
output = base_model.output
output = GlobalAveragePooling2D()(output)
#output = Dense(512, activation='relu')(output)
top_layers = Dense(133, activation='softmax')(output)
finetune_model = Model(inputs=base_model.input, outputs=top_layers)
finetune_model.summary()
finetune_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
steps = math.ceil(train_tensors.shape[0]/batch_size)
validation_steps = math.ceil(valid_tensors.shape[0]/batch_size)
print('Training Samples: {} Validation Samples: {} Batch Size: {} Steps: {}'.format(
train_tensors.shape[0], valid_tensors.shape[0], batch_size, steps))
epochs = 20
epochtimer = EpochTimer()
hist = finetune_model.fit_generator(
train_generator,
validation_data=validation_generator,
epochs=epochs,
steps_per_epoch=steps,
validation_steps=validation_steps,
callbacks=[epochtimer])
show_history_graph(hist)
top_model_file = 'saved_models/weights.best.{}.hdf5'.format('inceptionv3_top')
finetune_model.save(top_model_file)
predictions = [np.argmax(finetune_model.predict(np.expand_dims(feature, axis=0))) for feature in inception_test_tensors]
# report test accuracy
test_accuracy = 100*np.sum(np.array(predictions)==np.argmax(test_targets, axis=1))/len(predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
from keras.optimizers import SGD
# NB_IV3_LAYERS corresponds to the top 2 inception blocks in the inceptionv3 architecture
# Those should be finetuned.
NB_IV3_LAYERS_TO_FREEZE = 172
for layer in finetune_model.layers[:NB_IV3_LAYERS_TO_FREEZE]:
layer.trainable = False
for layer in finetune_model.layers[NB_IV3_LAYERS_TO_FREEZE:]:
layer.trainable = True
finetune_model.summary()
# Use a slow stable learning method to prevent introducing bad noise into the lower layers
finetune_model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy', metrics=['accuracy'])
steps = math.ceil(train_tensors.shape[0]/batch_size)
validation_steps = math.ceil(valid_tensors.shape[0]/batch_size)
print('Training Samples: {} Validation Samples: {} Batch Size: {} Steps: {}'.format(
train_tensors.shape[0], valid_tensors.shape[0], batch_size, steps))
epochs = 20
epochtimer = EpochTimer()
hist = finetune_model.fit_generator(
train_generator,
validation_data=validation_generator,
epochs=epochs,
steps_per_epoch=steps,
validation_steps=validation_steps,
callbacks=[epochtimer])
show_history_graph(hist)
finetune_model_file = 'saved_models/weights.best.{}.hdf5'.format('inceptionv3_finetune')
finetune_model.save(finetune_model_file)
predictions = [np.argmax(finetune_model.predict(np.expand_dims(feature, axis=0))) for feature in inception_test_tensors]
# report test accuracy
test_accuracy = 100*np.sum(np.array(predictions)==np.argmax(test_targets, axis=1))/len(predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
finetune_model_file = 'saved_models/weights.best.{}.hdf5'.format('inceptionv3_finetune')
finetune_model.load_weights(finetune_model_file)
from keras.applications.inception_v3 import InceptionV3
from keras.applications.inception_v3 import preprocess_input as preprocess_inception_input
def extract_InceptionV3(tensor):
return InceptionV3(weights='imagenet', include_top=False).predict(preprocess_inception_input(tensor))
### Function that takes a path to an image as input
### and returns the dog breed that is predicted by the finetuned model.
def breed_detector(img_path, selected_model, bottleneck=True, img_width=223, img_height=223):
tensor = path_to_tensor(img_path, img_width, img_height)
if bottleneck:
tensor = extract_InceptionV3(tensor)
else:
tensor = preprocess_inception_input(tensor)
predictions = selected_model.predict(tensor)
y_hat = np.argmax(predictions)
return dog_names[y_hat]
Results¶
With data augmentation and only training the top layers: Test accuracy: 85.8852%
After finetuning for 20 epochs with a very small learning rate and SGD optimizer the test accuracy went up to: 87.7990%. Based on the figures it is quite possible to achieve an even better learning rate with much longer training times.
It was really helpful and I learned a lot for potential future problems doing this additional exercise and would recommend to include it or at least some tutorial on how to do this.
# History graphs for the training of the top layers:
show_history_graph(hist)
# History graphs for the fine tuning of the
# NB_IV3_LAYERS which correspond to the top 2 inception blocks in the inceptionv3 architecture
show_history_graph(hist)